# Proof audit pass 2 — issue 24113

## Findings first

1. **Exact wide-sidebar bug is proven fixed on web.** S11–S15 show the base overlap and the head compact controls for all five fixtures in both themes. S16–S25 prove the changed container boundaries and full-width behavior. Independent S11-light rebuild has **zero changed pixels anywhere in the main pane** versus the submitted head image. The scope-correct affected web proof is sufficient.
2. **Original matrix is insufficient.** S26 is absent; R02 does not prove pointer dragging/nonzero animation; R03 does not establish the transient busy frame; R04 omits the required selection sequences and long label. S02/S04/S06/S08/S10 fail the original universal no-overlap criterion on both revisions. These unchanged mobile, refresh, and selection paths must not be conflated with the desktop issue.
3. **R01 persistence is overstated.** Both delivered videos end on the T3 loading splash after reload, not a settled selection. The action driver starts reload without waiting for the restored page. Starting a reload is not proof of persisted selection. The other recorded selections/key actions are useful regression evidence, but this requirement remains unproven.
4. **Electron remains explicitly unverified.** S26 has no native-shell images. ServerAlreadyRunningError is a capture blocker, not a passing shell check. Per the lead's supplied hard-stop policy, stop attempting this unavailable surface and list it in any eventual PR: “Electron native shell unverified; attempted launch failed with ServerAlreadyRunningError; web evidence does not establish native drag-region/control insets.” No PR was created or changed by this audit.
5. **Capture provenance is limited, not contradicted.** The final sanitized launch hint was added after capture. The reused backend actually has HOME=/home/ubuntu and unset XDG_CONFIG_HOME; its T3 state is the isolated worktree .t3. The lead reports zero host Grok/Pi usage files and absent Antigravity; submitted visible totals are synthetic Claude only. This is not proof that the original process was fully isolated. No leak was identified. The new env -i hint is source-verified, not runtime-verified here.

## Row statuses

PASS means the requested static layout proof is present, not that every unrelated interaction or surface was tested. REGRESSION FAIL means an unchanged preexisting condition violates the broader original matrix. PARTIAL never counts as PASS. L+D means both themes; all listed still rows have matched base/head images and both requested diff maps unless marked missing.

| Row | Evidence checked | Original matrix status | Affected-scope interpretation |
|---|---|---|---|
| S01 | C-D, 1440×900, sidebar256, L+D | PASS | Roomy baseline retained |
| S02 | C-D, 390×844, L+D | REGRESSION FAIL | Provider/metric overlap on both revisions; unchanged path |
| S03 | C-H, 1440×900, sidebar256, L+D | PASS | Roomy baseline retained |
| S04 | C-H, 390×844, L+D | REGRESSION FAIL | Same preexisting overlap |
| S05 | T-D, 1440×900, sidebar256, L+D | PASS | Roomy baseline retained |
| S06 | T-D, 390×844, L+D | REGRESSION FAIL | Same preexisting overlap |
| S07 | T-H, 1440×900, sidebar256, L+D | PASS | Roomy baseline retained |
| S08 | T-H, 390×844, L+D | REGRESSION FAIL | Same preexisting overlap |
| S09 | Limits, 1440×900, sidebar256, L+D | PASS | No span, period disabled |
| S10 | Limits, 390×844, L+D | REGRESSION FAIL | Same preexisting overlap |
| S11 | C-D, 1280×900, sidebar640, content600, L+D | PASS | Exact bug; independently rebuilt head light |
| S12 | C-H, same geometry, L+D | PASS | Exact bug |
| S13 | T-D, same geometry, L+D | PASS | Exact bug |
| S14 | T-H, same geometry, L+D | PASS | Exact bug |
| S15 | Limits, same geometry, L+D | PASS | Exact bug; disabled compact period retained |
| S16 | C-H, viewport1191×900, content895, light | PASS | Compact below boundary |
| S17 | C-H, viewport1192×900, content896, light | PASS | Expanded at boundary |
| S18 | C-H, viewport1193×900, content897, light | PASS | Expanded above boundary |
| S19 | Limits, content895, light | PASS | Compact, no span |
| S20 | Limits, content896, light | PASS | Expanded, disabled period, no span |
| S21 | Limits, content897, light | PASS | Expanded, disabled period, no span |
| S22 | C-H, viewport1351×900, content1279, collapsed, light | PASS | Date span absent |
| S23 | C-H, viewport1352×900, content1280, collapsed, light | PASS | Date span appears |
| S24 | C-H, viewport1353×900, content1281, collapsed, light | PASS | Date span remains |
| S25 | C-H, 1440×900, content1368, collapsed, L+D | PASS | Span and controls coexist |
| S26 | Native Electron, L+D | MISSING | Unavailable surface; disclose, do not infer from web |
| R01 | Both videos, action driver, captions, scripts | PARTIAL | Selections/keys exercised; reload persistence not visible; unrelated handlers unchanged |
| R02 | Both videos, width stops/collapse | PARTIAL | Changed branch transitions supported by stronger still geometry; CSS width writes and 0ms animation do not prove rail dragging/animated timing |
| R03 | Both videos, refresh/format actions | PARTIAL | Settled refresh/format flow present; transient busy state not established; unchanged event handlers/preferences |
| R04 | Both videos, menus/dismissal | PARTIAL | Long names, full selection sequences, zero-selection body/status case missing; programmatic desktop activation cannot prove pointer reachability |
| R05 | Both videos, opening and closing frames | PASS for overlay cycle | Responsive web overlay visibly opens then closes; preexisting header overlap remains; native mobile not tested |

## Revisions and code support

Issue [#24113](https://git.cafe/pingdotgg/t3code/issues/24113). Evidence declares base [eb1510d249](https://git.cafe/pingdotgg/t3code/commit/eb1510d249) and product head [2e7b0091fd](https://git.cafe/pingdotgg/t3code/commit/2e7b0091fd3b44a4a01912c055c3294fc09fc53e). Fixture at audit entry was [b26c26e20e](https://git.cafe/pingdotgg/t3code/commit/b26c26e20e4f85574be5fca8b62c1059ed8351e9); current final head is [dd03e05737](https://git.cafe/pingdotgg/t3code/commit/dd03e057376cc3c4d6c4b90d69e63165ead4572e). I verified product UsagePage is unchanged between product capture head and current HEAD. Current additions are the fixture script and knip configuration; repository status is clean. Historical image SHA attribution comes from the evidence manifest/capture account, not an embedded cryptographic screenshot attestation.

The product diff changes only UsagePage: viewport xl rules become @4xl/usage-header rules, the date span becomes @7xl/usage-header, the expanded group gets shrink-0, and WorkspacePageHeader establishes the named container. No refresh handlers, menu selection handlers, timestamp preference logic, native-mobile code, provider adapters or wire contracts changed. Runtime compiled rules include usage-header width >=56rem and >=80rem; root font is16px, matching896/1280. The extra72rem rule present in the stylesheet is not one of this diff's changed thresholds.

At base S11, breadcrumb extends x660–883.21875 while Cost occupies708.390625–761.515625: a direct intersection. Head S11 breadcrumb ends1061.5625 and Cost starts1073.5625, a12px gap. Across head S11–S15, provider/metric rectangles do not intersect; header height remains52px. The lead's separately supplied original pointer-drag repro is consistent with these measurements, but I did not need to substitute that recording for the submitted matrix artifacts.

The written differences are supported: S11–S15 expanded→compact, S17/S18/S20/S21 compact→expanded, and S23/S24/S25 date-span appearance follow the changed classes and measured geometry. Other nonzero bands include sidebar elapsed age, pointer/focus background, and hydration noise; these are disclosed rather than masked. An unchanged mobile overlap is not evidence of a new regression in this fix.

## Independent artifact checks

Read the local index plus coverage.md, differences.md, recordings.md, environment.md, artifacts.md, captures.json, records.json, pixel-stats.json, recording-driver.js, and build-proof.py. Inspected paired header contact sheets and sampled decoded frames from all10 delivered videos, including opening/closing R05 and terminal R01 frames.

All82 PNGs exist, form41 matched pairs, match declared light/dark backgrounds, and have matching viewport/sidebar/content-width geometry. Every metadata entry reports DPR2,16px root,UTC; all headers are52px; recorded document scrollWidth never exceeds viewport. Export dimensions differ from CSS viewport as disclosed: desktop preview downsampling and phone780×1688. No hidden resampling was added for this audit's pixel comparisons.

Independently recomputed **all41** absolute RGB maps, saturated zero×8 maps, threshold16 maps, and pixel counts from original PNGs: **zero mismatches**. Threshold means max RGB channel delta >16. Independently checked head breadcrumb-button/control intersections: only the ten phone row/theme combinations intersect, matching the disclosed preexisting issue.

All10 MP4s have video and audio tracks. The same narration source is used per before/after pair; scripts describe the driver's actions and explicitly call R04 partial. Narration is an overview synthesized after capture, not synchronized per click. Event t values start before setup/recording, so they are not direct MP4 seek timestamps. Desktop resize phases are letterboxed into a780×1688 recording canvas, reducing readable header detail. Captions/action source support intent; neither narration nor a caption alone proves completion. R01's held terminal loading frame is the concrete exception to the claimed visible persistence. No independent speech transcription was performed.

## One real-app rebuild

Opened my own preview tab with reuseExistingTab:false: tab_1h_718af28e-b901-447a-834c-af7f6f2e4896. No other agent tab operated. Paired normally with a fresh sandbox token, used the real /usage route and existing server. Set Cost/90d through the existing preference, light appearance,1280×900 viewport, expanded640px sidebar via its existing CSS variable. No product CSS, fixture source, or repository files were edited.

Executed the committed final fixture script against `/tmp/issue-24113-proof-audit/pass2/seed-check`, then compared its Claude JSONL with the running sandbox fixture: byte-identical. Reused that existing fixture/server rather than rewriting its settings or restarting it. Browser DPR2,UTC,16px root,header52px,content600px,main640px match S11. Visible totals remain$1.26 and252K tokens. This reproduces the original rendering environment; it does not demonstrate a launch under the new sanitized environment hint.

Rebuilt image and submitted head S11-light are both1280×900. Full-image pixel delta: **847/1,152,000 (0.07352%) at zero threshold;781 (0.06780%) at threshold16**. Header band has744/709 changed pixels; below-header103/72. Bounding box is[13,12,609,125], wholly inside sidebar. **Main pane x>=640:0 changed pixels**, including the complete Usage header/body. Differences are pointer background and elapsed label in the sidebar, not layout drift. Raw full-frame comparison had no masks/alignment/resampling; main-pane count is an additional localization measurement.

![Independent rebuilt head S11-light](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b9de658e8e0b6e37/rebuilt-head-S11-light.png)

[Raw rebuilt-versus-submitted pixel diff](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e5c9382261eae95d/rebuilt-S11-diff.png). [R01 head terminal frame showing loading splash](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/07111c76a27187e9/head-R01-final.png).

## Process and authorization record

Read-only audit with explicitly authorized browser/capture and scratch report/fixture execution. No product/fixture edits, checkout, push, comment, PR, or other worktree. No live T3 database was opened or modified; normal preview tools produced their app-managed browser artifacts. No server was started, stopped or restarted. Lead server remains available: web7771/PID2093610, backend15811/PID2093623, original session16025; listening owners and /proc cwd confirmed as this worktree's apps/web and apps/server. My preview tab remains on rebuilt Usage. Scratch image/video extraction processes do not own application servers.

## Friction

| FRICTION category | happened | avoidance |
|---|---|---|
| SCOPE | Original matrix demands unchanged mobile/menu/refresh behaviors | Separate exact affected web proof from broad regression coverage; retain missing statuses |
| RECORDING | R01 reload ends on splash; timestamps include pre-record setup | Wait for settled selected controls before stopping; record relative-to-record-start times |
| CAPTURE | Desktop video letterboxed into phone canvas; still exports downsampled | Prefer separate fixed-size recordings; retain CSS and exported dimensions |
| DATA | Captures and rebuild reused original unsanitized host env | Future reproduction must use committed env -i launch hint; do not retroactively claim it was used |
| DESKTOP | Native shell attempt failed with ServerAlreadyRunningError | Respect hard stop and explicitly disclose unavailable Electron surface in PR |
| POINTER | Phone overlap preexists; desktop menu activation was programmatic | Preserve failure evidence; don't claim universal pointer reachability |
| AUTH | My fresh tab required explicit pairing form submission | Use fresh sandbox token and wait for authenticated route before fixture setup |

Scope-correct conclusion: **SUFFICIENT for the affected web bug and container boundaries (S01/S03/S05/S07/S09/S11–S25); Electron S26 explicitly UNVERIFIED under the supplied unavailable-surface policy.** Missing unchanged interactions do not justify unbounded recapture or product scope expansion.

VERDICT INSUFFICIENT original matrix: S02/S04/S06/S08/S10 universal no-overlap; S26 missing; R01 persistence; R02 pointer/animation timing; R03 busy-state proof; R04 selection/long-label/status coverage.